Differential Privacy in Bayesian Networks for Synthetic Data

Abstract

With data-driven decision-making becoming increasingly prevalent, maintaining the confidentiality of protected information that resides in database systems is now an acute issue. Classical privacy-preserving methods, including Differential Privacy (DP), provide firm theoretical assurance by injecting noise that is calibrated into the answer that is output by each query. Nevertheless, such protection usually comes at the price of utility, at least where the level of noise is preset and independent of user background knowledge. At the same time, inference attacks are perilous by virtue of allowing adversaries to infer covert information by tracking a series of apparently innocuous queries. Existing inference-detecting mechanisms attempt to model and observe user knowledge but lack any effective control measures for probabilistically handling information leakage. This dissertation studies a hybrid solution that blends Differential Privacy with user-adapted knowledge modeling for adaptively regulating the level of privacy. With probabilistic graphical models, notably Bayesian networks, we want to predict what the user would infer from the response to queries. The level of noise due to DP is then fixed based on this estimated knowledge such that utility and privacy are traded off. Our approach is validated by an implementation prototype with a PostgreSQL database, populated with inference monitoring and DP protection. The result demonstrates utility growth with stronger privacy protection by varying protection with user knowledge.